A bearing imbalance fault diagnosis method based on generative data fusion

Abstract

This invention relates to a bearing imbalance fault diagnosis method based on generative data fusion, comprising: data preprocessing: acquiring a bearing vibration time-series signal dataset, the dataset including labeled minority class fault samples and majority class normal samples; normalizing the minority class fault samples; performing a first data augmentation on the minority class fault samples based on a Wasserstein generative adversarial network; performing a second data augmentation on the minority class fault samples using a clustering-based minority class sample synthesis technique; concatenating and fusing the samples from the first and second data augmentations with the same number of samples, and summing them with the majority class normal samples to obtain a fused dataset; training an attention-based convolutional neural network diagnostic model using the fused dataset; and performing fault diagnosis using the trained convolutional neural network diagnostic model. Compared with existing technologies, this invention maintains reliable accuracy even with a high dataset imbalance rate.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a bearing imbalance fault diagnosis method based on generated data fusion. Background Technology

[0002] Fault diagnosis is crucial for identifying the health status of mechanical equipment. Bearings, as core components of rotating machinery, are an indispensable part of many industrial machines. In recent years, a large number of industrial accidents have been caused by bearing damage; therefore, accurate and timely assessment and detection of bearing operational health is extremely important.

[0003] With the emergence of machine learning and deep learning methods, data-driven fault diagnosis methods have become a research hotspot in recent years. Methods such as support vector machines, fuzzy logic, and neural networks have been widely applied in fault diagnosis. Deep learning models can improve the diagnostic accuracy of various bearing datasets to a very high level. However, various deep learning models typically require a large number of data samples for training. If the training samples have an imbalanced class distribution, their performance will be affected. In actual industrial settings, bearings usually operate under normal conditions, and the number of fault samples available is far less than that of bearings operating normally. This often leads to a decrease in the accuracy of data-driven model diagnostic methods. Summary of the Invention

[0004] The purpose of this invention is to provide a bearing imbalance fault diagnosis method based on generated data fusion, thereby improving the fault diagnosis accuracy under conditions of sample data imbalance.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A bearing imbalance fault diagnosis method based on generated data fusion includes the following steps:

[0007] Data preprocessing: The bearing vibration time series signal dataset is obtained based on sliding window cutting and splicing processing. The dataset includes labeled minority class fault samples and majority class normal samples. The labeled minority class fault samples are normalized.

[0008] A data augmentation is performed on the normalized minority class fault samples based on the Wasserstein generative adversarial network.

[0009] Secondary data augmentation of minority fault samples using clustering-based minority sample synthesis technology;

[0010] The samples from the first and second data augmentations are spliced ​​and fused with the same number of samples, and then summarized with the normal samples of the majority class to obtain the fused dataset.

[0011] Model training: Train an attention-based convolutional neural network diagnostic model using a fused dataset;

[0012] Fault diagnosis: Fault diagnosis is performed using a trained convolutional neural network diagnostic model.

[0013] The normalization process is as follows:

[0014]

[0015] Where a represents the vibration signal points in the dataset, A is the set of a, and a' represents the normalized vibration signal points. The normalization process maps the original signal values ​​to the interval [0, 1] through max-min standardization.

[0016] The Wasserstein generative adversarial network consists of a generator and a discriminator. The generator comprises one linear transformation layer and four transposed convolutional layers, while the discriminator consists of four convolutional layers and one fully connected layer. The activation function of each layer is a ReLU function. The Wasserstein generative adversarial network uses the Wasserstein distance to measure the difference between two distributions and employs a gradient penalty mechanism to optimize the discriminator's loss function. The Wasserstein distance is:

[0017]

[0018] Where, p r The distribution of real data, p g The distribution of the data generated by the generator, ∏(P) r ,P g Let represent the set of all joint distributions formed by combining two distributions, γ represent each possible joint distribution, x, y refer to the samples obtained from γ, ||xy|| refers to the distance between the two distributions, and E represents the expected value of the distance between the samples.

[0019] The loss functions of the generator and discriminator in the Wasserstein generative adversarial network are determined based on Wasserstein distance, wherein the loss function of the generator is:

[0020]

[0021] Where, p z is the distribution of input random noise, z refers to the sample that follows the random noise distribution, E represents the mathematical expectation, and D(·) and G(·) represent the discriminator network and the generator network, respectively;

[0022] The loss function of the discriminator is:

[0023]

[0024] in, This is a gradient penalty mechanism, where λ represents the penalty coefficient. This refers to random interpolation sampling between samples generated by the generator and real samples. express The distribution that is satisfied, This represents the gradient of the discriminator. The Euclidean norm of the gradient is denoted as .

[0025] Wasserstein generative adversarial networks augment normalized fault samples using the following steps:

[0026] Normalized minority class fault samples are used as the training set and input into the generative adversarial network model, along with random noise z = (z 1 ,z 2 ,z 3 ,...,z m );

[0027] Train the generator G and initialize the parameters θ G Simulation data is generated based on the input random noise.

[0028] Based on the simulated data generated by generator G and the labeled real data, a discriminator D is trained, and the parameters ω of discriminator D are updated simultaneously. D The parameter ω D The update method is as follows:

[0029]

[0030]

[0031] Where, η D Let m represent the learning rate, and z represent the length of the generated signal. i x represents a noise sample. i Represents a real sample;

[0032] With the discriminator parameters fixed after training, the generator G's parameters θ are trained and updated. G The update method is as follows:

[0033]

[0034]

[0035] Where η G Indicates the learning rate. This represents the gradient of the generator;

[0036] The generator G and discriminator D are trained alternately in a loop until the data distribution generated by the generator is less than the true distribution, at which point it is determined that Nash equilibrium has been reached and the model training is completed.

[0037] A data augmentation of minority class fault samples is achieved based on the trained generator G.

[0038] The cluster-based minority class sample synthesis technique for secondary data augmentation of minority class fault samples employs a generation strategy based on an oversampling algorithm, including the following steps:

[0039] 3-1) Concatenate each minority class of fault samples with the majority class of normal samples to form a backup dataset;

[0040] 3-2) Randomly select k samples from the spare dataset as cluster centers for k-means clustering;

[0041] 3-3) Calculate the distance from each sample point to the center of each cluster based on Euclidean distance:

[0042]

[0043] Where, x i and y i Let n represent two sample points in space, and n represent the dimension of the sample. Each sample point selects the cluster center closest to it to form k clusters.

[0044] 3-4) For the k clusters formed, calculate the sample point in the cluster that has the smallest sum of distances to other sample points and take it as the new cluster center;

[0045] 3-5) Repeat steps 3-3) to 3-4) until the cluster centers no longer change, thus completing the clustering;

[0046] 3-6) Sort all clusters after clustering according to the number of minority class fault samples contained in the cluster from largest to smallest, select the top k clusters to complete the filtering step;

[0047] 3-7) Within the filtered cluster, select a sample x from the minority class of fault samples based on the sampling fold R. i and its R nearest neighbor samples x Ri ;

[0048] 3-8) sequentially at x i and x Ri New sample points are synthesized from these points using the following formula:

[0049] x R =x i +rand(0,1)×|xRi -x i |

[0050] Where rand(0,1) represents generating a random number between 0 and 1, x R This indicates a newly synthesized sample;

[0051] 3-9) Repeat steps 3-1) to 3-8) for each dataset formed by splicing minority class fault samples and majority class normal samples to obtain minority class fault samples with secondary data augmentation.

[0052] The convolutional neural network diagnostic model includes a convolutional neural network and an attention mechanism. The convolutional neural network extracts features from the input data, and its output features are used as input to the attention mechanism for information weight allocation. The attention mechanism outputs fault features containing important information markers, and the fault diagnosis classification result of the diagnostic model is obtained through a fully connected layer and a softmax function. The loss function of the convolutional neural network diagnostic model is the cross-entropy loss function.

[0053]

[0054] Where J(θ) is the loss value, θ is the training parameters of the model, and h k For the output category, f k (θ) represents the predicted probability value of the sample, and N is the number of fault categories.

[0055] The convolutional neural network extracts features based on convolutional layers and pooling layers. The calculation process of the convolutional layer is as follows:

[0056]

[0057] in Let f(·) be the j-th feature map of the l-th layer after convolution, f(·) denotes the activation function, and M be the number of input feature maps. Let * represent the i-th feature map of the (l-1)-th layer of the input, and * denote the convolution operation. For convolution kernel, This represents the bias parameter; the activation function f uses the ReLU activation function.

[0058]

[0059] The convolutional layer is immediately followed by a pooling layer. The pooling process is as follows:

[0060] x j l+1 =f[down(x j l )+b l+1 j ]

[0061] in Let j be the j-th feature map of layer l+1 after pooling, and down be the downsampling function. These are the bias parameters for layer l+1;

[0062] The input data undergoes feature extraction through two convolutional and pooling layers to obtain the output features F = [f1, f2, f3, ..., f n ], where n is the length of the feature sequence.

[0063] The attention mechanism is based on a fully connected layer to calculate the attention score s of the features. i :

[0064] s i =sigmoid(w a f i +b a )

[0065] Where w a These are the weight parameters of the attention layer, b a It is the bias parameter of the attention layer, f i Given the input features, sigmoid is the activation function, and its expression is:

[0066]

[0067] The feature sequence F = [f1, f2, f3, ..., f3] output by the convolutional neural network is... n After inputting into the fully connected layer, we obtain the attention score sequence S = [s1, s2, s3, ..., s...]. n ];

[0068] The attention weight parameters to be trained are calculated using the attention scores, and the weights of the features are redistributed based on these attention weight parameters:

[0069]

[0070] f i '=α i f i

[0071] Where α i f represents the attention weight parameters obtained during training. i F′ represents the feature containing important information after weight redistribution via the attention mechanism. The output feature sequence of the attention mechanism is represented as F′=[f1′,f2′,f3′,...,f n ′).

[0072] The fully connected layer flattens the fault features, which contain important information markers, into one dimension:

[0073]

[0074] Where y i This represents the output of the fully connected layer, where σ(·) is the activation function, and w f and b f This refers to the weight matrix and bias parameters between adjacent layers;

[0075] The fault diagnosis and classification results are obtained after the fully connected layer using the Softmax function.

[0076]

[0077] Where f k (y) represents the predicted value for the k-th output, y k This represents the k-th output of the fully connected layer.

[0078] Compared with the prior art, the present invention has the following beneficial effects:

[0079] (1) The data fusion process of this invention adopts two strategies for data generation: generative adversarial network and oversampling algorithm, which enriches the sample features in the generated data while maintaining the completeness of the sample features in the new dataset.

[0080] (2) The present invention uses a convolutional neural network with an attention mechanism, which can better judge the extracted feature information, increase the weight of key features, reduce the influence of unnecessary features, and improve the ability to extract effective features.

[0081] (3) The present invention can maintain a high fault diagnosis accuracy when the sample imbalance rate is high.

[0082] (4) The Wasserstein generative adversarial network optimizes the loss function of the model by introducing the Wasserstein distance, thereby improving the stability of the generative adversarial network during training and the quality of the generated samples.

[0083] (5) The sampling process of the oversampling algorithm does not require a complex calculation process, has a low computational load, and can effectively supplement the number of minority class samples while enriching the features of the dataset. Attached Figure Description

[0084] Figure 1 This is a flowchart of the method of the present invention;

[0085] Figure 2 This is a schematic diagram of the model structure of the present invention;

[0086] Figure 3 This is a schematic diagram of the attention mechanism. Detailed Implementation

[0087] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0088] A bearing imbalance fault diagnosis method based on generative data fusion, the method flow is as follows: Figure 1 As shown, the model structure and data processing procedure are as follows: Figure 2 As shown, the specific steps include:

[0089] 1) Data preprocessing

[0090] This embodiment selects the publicly available Case Western Reserve University bearing dataset as the original database. The selected data type is drive-end fault data with a sampling frequency of 12kHz. The fault types included in the data are divided into nine categories: inner rings, rolling elements, and outer rings with diameters of 0.07, 0.14, and 0.21 inches. Including the normal samples without faults, there are a total of 10 data samples.

[0091] To more closely approximate the sample imbalance situation in real working conditions, different numbers of samples were extracted from the original dataset through sliding window cutting and splicing to obtain a bearing vibration time series signal dataset. The dataset includes labeled minority class fault samples and majority class normal samples, as shown in Table 1.

[0092] Table 1 Bearing vibration time sequence signal dataset

[0093] Fault type Fault Dimensions Sample size Label normal 0 300 1 Inner ring fault 0.07 30 2 Inner ring fault 0.14 30 3 Inner ring fault 0.21 30 4 Sphere failure 0.07 30 5 Sphere failure 0.14 30 6 Sphere failure 0.21 30 7 Outer ring fault 0.07 30 8 Outer ring fault 0.14 30 9 Outer ring fault 0.21 30 10

[0094] The initial imbalanced dataset contains 300 normal samples and 30 samples for each of the other 9 fault classes; the balanced dataset used for testing contains 50 samples for each class.

[0095] Normalize the minority class of fault samples with labels;

[0096]

[0097] Where a represents the vibration signal points in the dataset, A is the set of a, and a' represents the normalized vibration signal points. The normalization process maps the original signal values ​​to the interval [0, 1] through max-min standardization.

[0098] 2) Perform a data augmentation on the normalized minority class fault samples based on the Wasserstein generative adversarial network;

[0099] The Wasserstein generative adversarial network consists of a generator and a discriminator. The generator comprises one linear transformation layer and four transposed convolutional layers, while the discriminator consists of four convolutional layers and one fully connected layer. Each layer uses the ReLU activation function. Compared to the original generative adversarial network model, the Wasserstein generative adversarial network uses the Wasserstein distance instead of the JS divergence to measure the difference between two distributions and employs a gradient penalty mechanism to optimize the discriminator's loss function. The Wasserstein distance is:

[0100]

[0101] Where, p r The distribution of real data, p g The distribution of the data generated by the generator, ∏(P) r ,P g Let represent the set of all joint distributions formed by combining two distributions, γ represent each possible joint distribution, x, y refer to the samples obtained from γ, ||xy|| refers to the distance between the two distributions, and E represents the expected value of the distance between the samples.

[0102] The loss functions of the generator and discriminator in the Wasserstein generative adversarial network are determined based on Wasserstein distance, wherein the loss function of the generator is:

[0103]

[0104] Where, p z is the distribution of input random noise, z refers to the sample that follows the random noise distribution, E represents the mathematical expectation, and D(·) and G(·) represent the discriminator network and the generator network, respectively;

[0105] The loss function of the discriminator is:

[0106]

[0107] in, This is a gradient penalty mechanism, where λ represents the penalty coefficient. This refers to random interpolation sampling between samples generated by the generator and real samples. express The distribution that is satisfied, This represents the gradient of the discriminator. The Euclidean norm of the gradient is denoted as .

[0108] Wasserstein generative adversarial networks augment normalized fault samples using the following steps:

[0109] Normalized minority class fault samples are used as the training set and input into the generative adversarial network model, along with random noise z = (z 1 ,z 2 ,z 3 ,...,z m );

[0110] Train the generator G and initialize the parameters θ G Simulation data is generated based on the input random noise.

[0111] Based on the simulated data generated by generator G and the labeled real data, a discriminator D is trained, and the parameters ω of discriminator D are updated simultaneously. D The parameter ω D The update method is as follows:

[0112]

[0113]

[0114] Where, η D Let m represent the learning rate, and z represent the length of the generated signal. i x represents a noise sample. i Represents a real sample;

[0115] With the discriminator parameters fixed after training, the generator G's parameters θ are trained and updated. G The update method is as follows:

[0116]

[0117]

[0118] Where η G Indicates the learning rate. This represents the gradient of the generator;

[0119] The generator G and discriminator D are trained alternately in a loop until the data distribution generated by the generator is less than the true distribution, at which point it is determined that Nash equilibrium has been reached and the model training is completed.

[0120] A data augmentation of minority class fault samples is achieved based on the trained generator G.

[0121] 3) Secondary data augmentation of minority fault samples using clustering-based minority sample synthesis technology;

[0122] Using a single generative network strategy for data generation can effectively augment the features contained in the original dataset. To ensure the diversity of generated features, oversampling algorithms are also introduced for data augmentation. By clustering imbalanced samples and then using minority class sample synthesis techniques for data augmentation, the generated data can be kept within a safe range, reducing the tendency of sample marginalization and maintaining the high quality of generated samples.

[0123] 3-1) Each of the 9 minority class fault samples is concatenated with the majority class normal samples to form a backup dataset;

[0124] 3-2) Randomly select k samples from the spare dataset as cluster centers for k-means clustering;

[0125] 3-3) Calculate the distance from each sample point to the center of each cluster based on Euclidean distance:

[0126]

[0127] Where, x i and y i Let n represent two sample points in space, and n represent the dimension of the sample. Each sample point selects the cluster center closest to it to form k clusters.

[0128] 3-4) For the k clusters formed, calculate the sample point in the cluster that has the smallest sum of distances to other sample points and take it as the new cluster center;

[0129] 3-5) Repeat steps 3-3) to 3-4) until the cluster centers no longer change, thus completing the clustering;

[0130] 3-6) Sort all clusters after clustering according to the number of minority class fault samples contained in the cluster from largest to smallest, select the top k clusters to complete the filtering step;

[0131] 3-7) Within the filtered cluster, select a sample x from the minority class of fault samples based on the sampling fold R. i and its R nearest neighbor samples x Ri ;

[0132] 3-8) sequentially at x i and x Ri New sample points are synthesized from these points using the following formula:

[0133] x R =x i +rand(0,1)×|x Ri -x i |

[0134] Where rand(0,1) represents generating a random number between 0 and 1, xR This indicates a newly synthesized sample;

[0135] 3-9) Repeat steps 3-1) to 3-8) for each dataset formed by splicing minority class fault samples and majority class normal samples to obtain minority class fault samples with secondary data augmentation.

[0136] 4) The samples from the first data augmentation and the samples from the second data augmentation are spliced ​​and merged with the same number of samples (i.e., a 1:1 ratio), and then summarized with the normal samples of the majority class to obtain a fused dataset. The number of each sample in the fused dataset is 300.

[0137] The minority class fault samples generated in steps 3) and 4) are concatenated with the same number of samples to generate a new fault sample dataset that is perfectly balanced and contains feature types emphasized by the two different generation strategies. This new fault sample dataset, along with the majority class normal samples, is then used as the training set and fed into an attention-based convolutional neural network diagnostic model.

[0138] 5) Model training: Train an attention-based convolutional neural network diagnostic model using the fused dataset;

[0139] The convolutional neural network diagnostic model includes a convolutional neural network and an attention mechanism. The convolutional neural network extracts features from the input data, and its output features are used as input to the attention mechanism for information weight allocation. The attention mechanism outputs fault features containing important information markers, and the fault diagnosis classification result of the diagnostic model is obtained through a fully connected layer and a softmax function. The loss function of the convolutional neural network diagnostic model is the cross-entropy loss function.

[0140]

[0141] Where J(θ) is the loss value, θ is the training parameters of the model, and h k For the output category, f k (θ) represents the predicted probability value of the sample, and N is the number of fault categories.

[0142] The convolutional neural network extracts features based on convolutional layers and pooling layers. The calculation process of the convolutional layer is as follows:

[0143]

[0144] in Let f(·) be the j-th feature map of the l-th layer after convolution, f(·) denotes the activation function, and M be the number of input feature maps. Let * represent the i-th feature map of the (l-1)-th layer of the input, and * denote the convolution operation. For convolution kernel, This represents the bias parameter; the activation function f uses the ReLU activation function.

[0145]

[0146] The convolutional layer is immediately followed by a pooling layer. The pooling process is as follows:

[0147] x j l+1 =f[down(x j l )+b l+1 j ]

[0148] in Let j be the j-th feature map of layer l+1 after pooling, and down be the downsampling function. These are the bias parameters for layer l+1;

[0149] The input data undergoes feature extraction through two convolutional and pooling layers to obtain the output features F = [f1, f2, f3, ..., f n ], where n is the length of the feature sequence.

[0150] The attention mechanism is based on a fully connected layer to calculate the attention score s of the features. i :

[0151] s i =sigmoid(w a f i +b a )

[0152] Where w a These are the weight parameters of the attention layer, b a It is the bias parameter of the attention layer, f i Given the input features, sigmoid is the activation function, and its expression is:

[0153]

[0154] The feature sequence F = [f1, f2, f3, ..., f3] output by the convolutional neural network is... n After inputting into the fully connected layer, we obtain the attention score sequence S = [s1, s2, s3, ..., s...]. n ];

[0155] The attention weight parameters to be trained are calculated using the attention scores, and the weights of the features are redistributed based on these attention weight parameters:

[0156]

[0157] f i'=α i f i

[0158] Where α i f represents the attention weight parameters obtained during training. i F′ represents the feature containing important information after weight redistribution via the attention mechanism. The output feature sequence of the attention mechanism is represented as F′=[f1′,f2′,f3′,...,f n A schematic diagram of the attention mechanism is shown below. Figure 3 As shown.

[0159] The fully connected layer flattens the fault features, which contain important information markers, into one dimension:

[0160]

[0161] Where y i This represents the output of the fully connected layer, where σ(·) is the activation function, and w f and b f This refers to the weight matrix and bias parameters between adjacent layers;

[0162] The fault diagnosis and classification results are obtained after the fully connected layer using the Softmax function.

[0163]

[0164] Where f k (y) represents the predicted value for the k-th output, y k This represents the k-th output of the fully connected layer.

[0165] 6) Fault diagnosis: Fault diagnosis is performed using the trained convolutional neural network diagnostic model.

[0166] The fault diagnosis results of this embodiment are shown in Table 2.

[0167] Table 2 Accuracy of Fault Diagnosis Classification

[0168]

[0169] It can be seen that when there is data imbalance, the recognition accuracy of existing methods drops significantly as the imbalance intensifies. However, the classification accuracy of the fault diagnosis method based on generative data fusion proposed in this invention is greater than that of the two generative methods used individually, and far greater than the recognition accuracy of existing methods.

[0170] Table 3 shows a comparison of the accuracy improvement effect of the method described in this invention and the enhanced GAN. It can be clearly seen from the table that, compared with the enhanced GAN, the method proposed in this invention has a higher accuracy improvement under the two imbalance ratios of 1:10 and 1:20.

[0171] Table 3 Comparison of the improvement in recognition accuracy between the method of the present invention and enhanced GAN.

[0172]

[0173] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A bearing imbalance fault diagnosis method based on generated data fusion, characterized in that, The method comprises the following steps: Data preprocessing: based on the sliding window cutting and splicing processing, a bearing vibration time series signal dataset is obtained, the dataset comprises a few class fault sample with a label and a majority class normal sample, and the few class fault sample with a label is subjected to normalization processing; The normalized few class fault sample is subjected to data expansion once based on a Wasserstein generative adversarial network; The few class fault sample is subjected to secondary data expansion based on a clustering-based few class sample synthesis technology; The once data-expanded sample and the twice data-expanded sample are spliced and fused according to the same sample quantity, and are integrated with the majority class normal sample to obtain a fusion dataset; Model training: a convolutional neural network diagnosis model based on an attention mechanism is trained by using the fusion dataset; Fault diagnosis: the trained convolutional neural network diagnosis model is used for fault diagnosis. The secondary data expansion of the few class fault sample based on the clustering-based few class sample synthesis technology adopts a generation strategy based on an oversampling algorithm, and comprises the following steps: 3-1) each type of few class fault sample is spliced with the majority class normal sample to form a standby dataset; 3-2) k samples in the standby dataset sample are randomly selected as cluster centers of k-means clustering; 3-3) the distance of each sample point to each cluster center is calculated based on the Euclidean distance: wherein, and denote two sample points in space, denotes the dimension of the sample, each sample point chooses the cluster center closest to it to form k clusters; 3-4) For the formed k cluster, calculate the sample point with the minimum sum of distances to other sample points in the cluster as the new cluster center of the cluster; 3-5) steps 3-3) to 3-4) are repeated until the cluster center is no longer changed, and the clustering is completed; 3-6) Sort all clusters completed according to the number of minority class fault samples contained in the cluster from large to small, select the top k cluster, complete the filtering step; 3-7) In the filtered clusters, based on the sampling rate R selecting one sample from the minority class fault samples and its adjacent R nearest neighbor samples ; 3-8) Synthesizing new sample points between and in turn, the synthesis formula is as follows: wherein represents a random number between 0 and 1, represents a synthesized new sample; 3-9) the steps 3-1) to 3-8) are repeated for the dataset spliced from each type of few class fault sample and the majority class normal sample, and the few class fault sample subjected to secondary data expansion is obtained.

2. The bearing imbalance fault diagnosis method based on generated data fusion according to claim 1, characterized in that, The normalization processing is: wherein, a representing the vibration signal points in the data set, A is a a set of a’ denotes the normalized vibration signal points, the normalization process maps the original signal values to the interval [0, 1] by min-max normalization.

3. The bearing imbalance fault diagnosis method based on generated data fusion according to claim 1, characterized in that, The Wasserstein generative adversarial network comprises a generator and a discriminator, wherein the generator comprises a linear transformation layer and four transpose convolution layers, the discriminator comprises four convolution layers and a fully connected layer, the activation function of each layer is a Relu function, the Wasserstein generative adversarial network uses a Wasserstein distance to measure the difference between two distributions, and adopts a gradient penalty mechanism to optimize the loss function of the discriminator, wherein the Wasserstein distance is: wherein, denotes the distribution of real data, denotes the distribution of data generated by the generator, denotes the set of all joint distributions of the two distributions combined, denotes each of the possible joint distributions, denotes the sample obtained in denotes the distance between the two distributions, denotes the expected value of the distance for the pair of samples.

4. The bearing imbalance fault diagnosis method based on generated data fusion according to claim 3, characterized in that, The loss functions of the generator and the discriminator of the Wasserstein generative adversarial network are determined based on the Wasserstein distance, wherein the loss function of the generator is: wherein, is a distribution of input random noise, refers to a sample subject to the distribution of random noise, denotes a mathematical expectation, and denote a discriminator network and a generator network, respectively; The loss function of the discriminator is: wherein, is a gradient penalty mechanism, represents a penalty coefficient, refers to a random interpolation sample between the sample generated by the generator and the real sample, denotes a distribution satisfied by, denotes the gradient of the discriminator, denotes the Euclidean norm of the gradient.

5. The bearing imbalance fault diagnosis method based on generated data fusion according to claim 4, characterized in that, The data expansion of the normalized fault sample by the Wasserstein generative adversarial network comprises the following steps: The normalized minority fault samples are input into the generative adversarial network model as a training set, and random noise is also input ; Training the generator G and initializing the parameters , generating simulated data from input random noise ; According to the simulation data generated by the generator G and the real data with labels, the discriminator D is trained, and the parameters of the discriminator D are updated , the updating manner of the parameters is that wherein, denotes a learning rate, denotes a length of the generated signal, denotes a noise sample, denotes a real sample; fix the trained discriminator parameters, train and update the parameters of the generator G, the parameters of the generator G are updated in the following manner: wherein denotes the learning rate, denotes the gradient of the generator; The generator G and the discriminator D are subjected to alternating cyclic training until the data distribution generated by the generator is less than a set threshold value, and then it is determined that the Nash equilibrium is reached, and the model training is completed; The few class fault sample is subjected to data expansion once based on the trained generator G.

6. The bearing imbalance fault diagnosis method based on generated data fusion according to claim 1, characterized in that, The convolutional neural network diagnosis model comprises a convolutional neural network and an attention mechanism, the convolutional neural network extracts features from input data, the output features are used as input of the attention mechanism to perform weight distribution of information, the attention mechanism outputs fault features marked with important information, and a fault diagnosis classification result of the diagnosis model is obtained through a full connection layer and a Softmax function, and a loss function of the convolutional neural network diagnosis model is a cross-entropy loss function: wherein is a loss value, is a training parameter of the model, is an output class, denotes a predicted probability value of a sample, N is a number of failure classes.

7. The bearing imbalance fault diagnosis method based on generated data fusion according to claim 6, characterized in that, The convolutional neural network extracts features based on a convolutional layer and a pooling layer, wherein a calculation process of the convolutional layer is as follows: where is the th feature map of the layer after convolution is the th feature map of the layer after convolution is the activation function is the number of input feature maps is the th input feature map of the layer is the th feature map of the layer after convolution is the th feature map of the layer after convolution is the convolution operation is the convolution kernel is the bias parameter; the activation function ReLU activation function is adopted:​ A pooling layer is arranged immediately after the convolutional layer, and a process of the pooling is as follows: wherein is the pooling function, the first feature map of the layer, is the down-sampling function, is the bias parameter of the layer; The input data is subjected to feature extraction by two convolutional layers and pooling layers to obtain features of each channel output , is the length of the feature sequence.

8. The bearing imbalance fault diagnosis method based on generated data fusion according to claim 6, characterized in that, The attention mechanism calculates an attention score of the feature based on a fully connected layer : wherein is a weight parameter of the attention layer, is a bias parameter of the attention layer, is an input feature, sigmoid is an activation function, and the expression of which is: The feature sequence output by the convolutional neural network is The attention score sequence is obtained after inputting the full connection layer ; An attention score is used to calculate an attention weight parameter to be trained, and the features are re-distributed based on the attention weight parameter: wherein is the attention weight parameter obtained by training, is the feature containing important information marked after weight redistribution by the attention mechanism, and the output feature sequence of the attention mechanism is represented as .

9. The bearing imbalance fault diagnosis method based on generated data fusion according to claim 8, characterized in that, A full connection layer is used to flatten the fault features marked with important information into one dimension: wherein denotes the output of the fully connected layer, is an activation function, and is a weight matrix between adjacent layers and a bias parameter; A Softmax function is used to obtain a fault diagnosis classification result after the full connection layer: wherein represents the predicted value for the th output, represents the fully connected layer th output.

Images